Anthropic's September opened loud: a new flagship model, a big price lever, and a claim to the top of the coding benchmarks.
If you run agents through Codex and ChatGPT Work, the reasonable question is whether the math just changed under you. Here is the September 2026 comparison, with the parts vendors leave out.
The 45% figure is real but conditional: base Claude pricing is unchanged, cache reads dropped 75%, and Anthropic says the combination saves around 25% on typical workloads and up to around 45% on heavily agentic ones, on the metered API. Claude Code subscription users still lose 17% of weekly capacity on September 14. If you pay per token for agent workloads, test the new pricing. If you pay a flat subscription on either platform, nothing about your bill changed this week.
What Anthropic Actually Launched
Two models shipped on September 1, per Anthropic's announcement.
Claude Fable 5.1 is the new flagship for coding, knowledge work, and long-running problem-solving, available everywhere Claude runs: the API, Claude Code, claude.ai, and the major cloud platforms. Anthropic says it reaches results similar to or better than Fable 5 at lower effort settings, which is where much of the cost saving comes from, and that it fixes root causes in code instead of taking shortcuts.
Claude Mythos 5.1 is the same model with fewer restrictions, limited to vetted cybersecurity and life sciences professionals in Anthropic's trusted access programs. Unless your business is in one of those fields and enrolled, Fable 5.1 is the one that concerns you.
Two supporting claims worth noting with attribution. Anthropic says its newest safeguards block 60% fewer cybersecurity false positives, which matters if security scanning noise has been slowing your team's agent work. And MacRumors reports the model outperforms Fable 5, Opus 5, and OpenAI's GPT-5.6 Sol across multiple benchmarks. Benchmark leads change hands every quarter; treat that as this month's snapshot rather than a settled ranking.
Where the 45% Comes From
The headline number is a pricing mechanic, and understanding it tells you whether it applies to your business.
Base pricing did not move: $10 per million input tokens and $50 per million output tokens, the same as Fable 5. What dropped is the price of cache reads, down 75% to $0.25 per million tokens.
A cache read is what happens when the model re-reads context it has already seen: your project files, your instructions, the conversation so far. A chat assistant does this occasionally. An agent does it constantly, because every step of a multi-step task re-loads the same working context. That is why Anthropic's math lands where it does: around 25% savings for typical workloads, up to around 45% for complex coding and heavily agentic tasks.
The more agentic your workload, the more of your bill was cache reads, and the more of this cut you actually collect.
The Same Company Is Also Cutting Capacity
Hold the price cut next to the other September date on Anthropic's calendar.
On September 14, Claude Code subscription plans lose 17% of their current weekly usage capacity, a change Anthropic confirmed after its original announcement drew pushback. We covered the full Q4 capacity math when it was announced.
These two moves are not contradictory. They hit different billing surfaces. The API, where you pay per token, got cheaper for agent work. The subscription, where you pay flat and draw against weekly limits, got smaller. A cynic would say the pricing lever moved where Anthropic wins enterprise API volume, while the flat-rate plan most owners actually use tightened. Whatever the intent, the practical sorting is simple: metered users won this week, subscription users did not.
The September 2026 Comparison
| Question | Codex (OpenAI) | Claude (Anthropic) |
|---|---|---|
| How most owners pay | ChatGPT subscription tiers (Plus, Pro, Business) | Subscription plans, or metered API for custom agent workloads |
| What just changed on cost | OpenAI claims 10-50% more completed work per quota after August harness fixes | Cache reads down 75%; Anthropic says up to ~45% lower effective cost for agentic API work |
| Subscription capacity direction | Raised baselines plus periodic one-time boosts | Weekly capacity down 17% from current levels on September 14 |
| New model this month | GPT-5.6 Sol is the current flagship | Fable 5.1, which MacRumors reports leads it on multiple benchmarks this month |
| Ease of setup for a non-technical owner | Strong if you already live in ChatGPT: same account, connectors, and workspace | Comparable products exist; switching means new accounts, new connectors, new habits |
| Who should re-run the math now | Nobody urgently; your platform got more efficient in August | Anyone paying per token for agent workloads, especially heavy ones |
Context for this table: the Q4 capacity math, what OpenAI's harness fixes changed, and what businesses actually pay for per Ramp's spend data.
What to Do With This, by Situation
You run agents through ChatGPT and Codex subscriptions. Nothing about your bill changed. Your platform's August efficiency fixes were your version of a price cut. File the Claude news as a data point and keep operating.
You pay per token for agent workloads, on either platform. This is your trigger to re-run the math. Take one real workload, estimate what share of it is repeated context, and price it under the new cache rates. For heavy agent work, up to 45% is the difference between a tool that costs like software and one that costs like a contractor.
You are on Claude Code subscription plans. The date that matters for you is September 14, and the move is capacity planning, covered in the capacity post: audit your weekly usage, pre-run heavy batches, and decide which tasks are portable.
You are choosing a platform from scratch. Do not choose on this week's headline from either vendor. Run the same real workflow on both for two weeks, then compare three numbers: completed work per dollar, completed work per hour of your attention, and how often you had to intervene. The platform that wins your workload is the right one, whatever the benchmarks say that month.
The Pattern Worth Noticing
Step back from the individual numbers and September looks like this: both vendors are now competing on the cost of completed agent work, not on chat quality. OpenAI spent August making Codex waste less quota on overhead. Anthropic opened September making repeated context nearly free on its API.
That competition is good for you, and it rewards one specific habit: measuring your own cost per completed task. Owners who know that number can collect each round of savings as it lands. Owners who only know their subscription price will keep reading headlines like this one and wondering whether they should feel something.
Frequently Asked Questions
What did Anthropic launch on September 1, 2026?
Claude Fable 5.1, its newest flagship model for coding and long-running agent work, plus Claude Mythos 5.1, the same model with fewer restrictions that is limited to vetted cybersecurity and life sciences professionals in trusted access programs. Fable 5.1 is available everywhere Claude runs: the API, Claude Code, claude.ai, and the major cloud platforms.
Is Claude really 45% cheaper now?
Up to, and only for certain work. Base API pricing is unchanged at $10 per million input tokens and $50 per million output tokens. What dropped is cache-read pricing, down 75% to $0.25 per million tokens. Anthropic says that works out to roughly 25% lower costs for typical workloads and up to around 45% lower for complex coding and highly agentic tasks, because agent workloads re-read the same context constantly.
Does this cancel the September 14 Claude Code limit cut?
No. These are two different billing surfaces. The cache-read cut lowers per-token costs on the API, where usage is metered. The September 14 change reduces weekly usage capacity on Claude Code subscription plans by 17% from current levels. If you pay a flat subscription, the limit change is the one you will feel.
Should I move my business from Codex to Claude over this?
Not on a headline. Price cuts and capability claims change quarterly on both sides. The durable habit is a quarterly calibration test: run one real workflow on the other platform, compare completed work per dollar and per hour of your attention, and switch only when the gap is large and repeatable.